Add Optional Agentic Retrieval Support to AI Search #705
Description
We need AI Search to support agentic retrieval workflows (multi-step planning, iterative query refinement, and tool-driven retrieval). Today, AI Search assumes a single-shot or linear retrieval pattern, which limits more advanced agents that need to reason over multiple retrieval steps.
Proposed feature
Add optional “agentic retrieval” support aligned with the public preview described here:
https://learn.microsoft.com/en-us/azure/search/agentic-retrieval-overview?tabs=quickstarts
Key capabilities that would help:
- Multi-step retrieval orchestration (plan → retrieve → refine → retrieve)
- Built-in support for query rewrite and decomposition
- Native controls for tool selection, reranking strategy, and retrieval depth
- Traceability/telemetry for each retrieval step (for debugging and evaluation)
Important constraint
This should be opt-in and must not impact deployments in other Azure clouds, including Azure Government, Azure China, and Azure Germany. If agentic retrieval is unavailable in those clouds, the system should gracefully fall back to standard retrieval without behavior changes.
Why this matters
- Enables more accurate, context-aware responses for complex user questions
- Reduces application-side custom orchestration code
- Aligns AI Search with modern agent frameworks and best practices